Recurrent Relational Networks

By Rasmus Berg Palm, Ulrich Paquet, Ole Winther. 2017

168 citations recorded by the source corpus.

This paper is concerned with learning to solve tasks that require a chain of interdependent steps of relational inference, like answering complex questions about the relationships between objects, or solving puzzles where the smaller elements of a solution mutually constrain each other. We introduce the recurrent relational network, a general purpose module that operates on a graph representation of objects. As a generalization of Santoro et al. [2017]'s relational network, it can augment any neural network model with the capacity to do many-step relational reasoning. We achieve state of the art results on the bAbI textual question-answering dataset with the recurrent relational network, consistently solving 20/20 tasks. As bAbI is not particularly challenging from a relational reasoning point of view, we introduce Pretty-CLEVR, a new diagnostic dataset for relational reasoning. In the Pretty-CLEVR set-up, we can vary the question to control for the number of relational reasoning steps that are required to obtain the answer. Using Pretty-CLEVR, we probe the limitations of multi-layer perceptrons, relational and recurrent relational networks. Finally, we show how recurrent relational networks can learn to solve Sudoku puzzles from supervised training data, a challenging task requiring upwards of 64 steps of relational reasoning. We achieve state-of-the-art results amongst comparable methods by solving 96.6% of the hardest Sudoku puzzles.

Subfields: Computer Vision and Pattern Recognition, Model Architectures, Graph Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Model Architectures

In Graph Learning

Citations in the corpus

Reconciling deep learning with symbolic artificial intelligence: representing objects and relations (2019, background): “A network architecture incorporating stacked or recurrent relation or self-attention modules could be thought of as carrying out a series of logic-like reasoning steps [42,51].”
On the Binding Problem in Artificial Neural Networks (2020, background): “, 2020), visual question answering (Santoro et al., 2017; Palm et al., 2018), abstract visual reasoning (Andreas, 2019), natural language processing (Tai et al.”
Recurrent Event Network for Reasoning over Temporal Knowledge Graphs (2019, background): “GN [19, 2] and RRN [17] update node representations by a message passing scheme between time stamps.”
PARN: Position-Aware Relation Networks for Few-Shot Learning (2019, background): “[23] introduced relation networks (RNs) for few-shot learning, which are a kind of architectures for relational reasoning and successfully applied in visual question answering tasks [17, 20, 30].”
Combinatorial optimization and reasoning with graph neural networks (2021, methodology): “These include using the encode-process-decode paradigm [Hamrick et al. , 2018], favoring max -aggregation, or leveraging strong supervision [Palm et al. , 2017].”
SATNet: Bridging deep learning and logical reasoning using a differentiable satisfiability solver (2019, background): “We also show that we can learn to play 9 × 9 Sudoku, a problem that is challenging for modern neural network architectures (Palm et al., 2017).”
View-GCN: View-Based Graph Convolutional Network for 3D Shape Analysis (2020, unknown): “We deﬁne message from node v i to v j as a pairwise relation [38, 45]: where f li ∈ R d is feature of i -th node, i.e ., the i -th row of F l , [ · , · ] denotes the concatenation of two vectors.”
Neural Machine Translation: A Review (2019, background): “(2018) is to resort to (recursive) relation networks (Santoro et al., 2017; Palm et al., 2018) which repeatedly aggregate pairwise relations between words in the sentence.”
GEAR: Graph-based Evidence Aggregating and Reasoning for Fact Verification (2019, methodology): “Inspired by recent work on semi-supervised graph learning and relational reasoning (Kipf and Welling, 2017; Velickovic et al., 2018; Palm et al., 2018), we propose an evidence reasoning network (ERNet) to propagate information among the evidence nodes.”
Neural Learning of One-of-Many Solutions for Combinatorial Problems in Structured Output Spaces (2020, methodology): “Our preliminary analysis of a state-of-the-art neural Sudoku solver [Palm et al., 2018]2, which trains and tests on instances with single solutions, showed that it achieves a high accuracy of 96% on instances with single solution, but the accuracy drops to less than 25%, when tested on inputs that h”
Disentangling Physical Dynamics From Unknown Factors for Unsupervised Video Prediction (2020, background): “Relational reasoning, often implemented with graphs [2, 26, 53, 45, 59], can account for basic physical laws, e.”
Fast and flexible design of novel proteins using graph neural networks (2019, methodology): “As there had been little previous work in using neural networks to solve CSPs ( 24 , 25 ), we first had to devise a network architecture that would be well-suited for this problem.”
Recurrent Event Network: Autoregressive Structure Inference over Temporal Knowledge Graphs (2019, methodology): “. Recurrent Graph Neural Models. There have been some studies on recurrent graph neural models for sequential or temporal graph-structured data (Sanchez-Gonzalez et al., 2018; Battaglia et al., 2018; Palm et al., 2018; Seo et al., 2017; Pareja et al., 2019). These methods adopt a messagepassing fram”
Graph Neural Networks: A Review of Methods and Applications (2018, methodology): “ce LSTM to solve the multi-hop reading comprehension problem. Another important direction is relational reasoning, relational networks [96], interaction networks [4] and recurrent relational networks [118] are proposed to solve the relational reasoning task based on text. The works cited above are n”